WebApp-Based Digital Modulation Waveform Generator for Automatic Modulation Classification in Wireless Communication Systems

To alleviate the issue of data scarcity in the context of communication systems, a viable solution is the capacity to generate realistic wireless signal constellation diagrams with Generative Adversarial Networks (GANs). While it was possible to train stable GANs for multi-class (20 classes) image data for some time, the situation where it is launched with strict calculation resources such as only using the CPU and High Performance Storage(HPS) in a small amount, and is trained for a large number of classes (260 classes) has been difficult until recently. In this paper, we propose two lightweight GANs, an improved vanilla generator-dominant GAN (G:D ratio 3.28) and an improved vanilla GAN (G:D ratio 1.43), along with a novel hybrid GAN model with the incorporation of DCGAN components and class conditioning: a Conditional Least Squares GAN (LSGAN Hybrid, G:D ratio 2.0) is made. All models were trained on 52,000 constellation images with four successive epochs (5, 10, 15, 20) using a standard Intel Core i7 CPU. The LSGAN Hybrid achieved outstanding training stability with monotonically decreasing loss throughout training, and with the smallest epoch-to-epoch correlation variance (σ = 0.0073) with only 720K parameters, which is more than 55% reduction from its vanilla counterpart. All architectures had 100% discriminative accuracy (ACC) when tested against the test dataset; the vanilla models, however, were suffering from serious discriminator dominance, with the ratio reaching up to 18.57 for the 10th epoch, which reflects the occurrence of the gradient becoming saturated. SGD statistical distribution analyses (t-tests) showed that the LSGAN was the only algorithm to make significant distributional changes at subsequent epochs (p